This sounds like it would be doable with a hobbyist setup. You can scan the room with a single depth image from Kinect (or two or three if you have to tilt it), and you have a depth map with which to distort an image, which you then send to a bog-standard projector.
I wonder what kind of algorithm one would use for the distortion. They do the same thing with projections on buildings, and I've always wondered.
Edit: I am wondering how this looks in real life, though. Those Johnny Chung-Lee Wiimote videos looked amazing too, but when you see it in real life (and with two eyes, rather than through one camera lens) the effect is much less impressive. I'm thinking the same might apply to stuff like, say, the depth effect with the snowflakes.